Autonomous Lab

Commissioned research · human-supervised

Name a question. Get back real research.

Bring an idea — or just a field. We survey the literature, gather datasets and tools, form and test hypotheses, and run the loop under a research scientist's supervision. What lands in your inbox is a paper, not a chat log.

Preview a run See how it works

Live preview

Watch a run begin.

Type a field or pick one below. This is a scripted preview of how a commissioned run starts — the literature surfacing, the plan branching, and the checkpoint where you and a scientist sign off on direction before any compute runs.

try:
RUN ---- · "…"
IDLE
// no active run name a field above to preview how a run begins

What we do

A commissioned research service, not a self-serve tool.

Autonomous Lab studies recent conference papers, extracts research gaps, and turns a field or a loose idea into a scoped, testable question. A named research scientist is attached to every run from scope to sign-off — you set the direction and review the findings; the scientist and pipeline handle execution.

20,922 NeurIPS, CVPR & ICLR papers tracked
41 avg. core papers cited per deliverable
4 / 2 pipeline stages, human checkpoints
100% findings signed by a named scientist

How it works

Four stages. Two don't finish without a human.

Scope and Validate are gated. A run can't spend compute on a badly-formed question, and it can't ship a finding a scientist hasn't checked against the evidence.

01

Scope

Turns your field or loose idea into a concrete, testable question — with any constraints you set.

◆ Checkpoint

Direction review

You and the assigned scientist approve or rewrite the question before any compute runs. Narrow it, redirect it, or call it off.

02

Retrieve

Surveys the literature, pulls datasets, benchmarks and tools, and normalizes everything into one workspace.

03

Investigate

Registers competing hypotheses and runs the experiments that discriminate between them — with seeds and logs.

04

Validate

Checks results against the literature, quantifies uncertainty, and assembles the deliverable.

◆ Findings validation

Scientist verifies every claim & signs the log.

Human in the loop

A scientist owns every run.

Trust is the hard part, so we don't automate it away. A named research scientist is attached to your run from scope to sign-off — and they hold a kill switch the entire time.

  • Before

    Direction review

    You and the assigned scientist approve or rewrite the scoped question and plan before a single GPU-hour is spent.

  • During

    Interruptible

    The scientist watches the run live and can inject constraints, exclude sources, or force a different method mid-stream.

  • After

    Findings validation

    Every claim is traced to its evidence. The scientist checks it against raw data, annotates, then signs — or sends it back.

  • Always

    On the record

    The deliverable carries the reviewer's name and the full decision log. No anonymous automation, no black box.

RUN a7f3 · review action

hypothesis

H2 · additive stabilizes SEI across all temperatures

Rejected by reviewer
"The 2023 dataset this leans on was retracted in Feb 2026. Re-run against the replication set before making any temperature-independence claim."

— reviewing scientist · 14 Jul 2026

What you get

A deliverable, not a chat transcript.

Every commission ships the same three things: findings you can defend, the evidence graph behind them, and a method log any reviewer can audit.

FINDING 03 confidence: moderate (0.71)

Vinylene-carbonate additive suppresses SEI growth above 45 °C, but the effect reverses below 10 °C — a temperature dependence absent from the three most-cited studies of the additive.

supports [02] Nakamura & Silva · Cell Reports 2023
supports [05] Petrov & Chen · J. Chem. Phys. 2024
counters [01] Zhang & Okafor · Nat. Methods 2022
RS
reviewing scientist · annotation

Verified against the raw cycling data, not just the reported means. Marked tentative until a second cell chemistry reproduces the crossover; softened the original claim's wording.

Fields we run in

If the literature exists and the question is answerable from data, it's in scope.

Track record

Deliverables from these runs have cleared peer review.

Work that began as a commissioned run has gone on to be accepted at top venues, with our clients as authors. Acceptance notices below — author, title and program-committee details are redacted at our clients' request.

venues: ICLR ICML NeurIPS CVPR AAAI ACL
ICLR
program-chairs@iclr.cc
22 Jan 2026 · 04:11 UTC
ACCEPTED
subject
[ICLR 2026] Decision: Accept (Poster)
redacted for author privacy
ICML
icml-decisions@icml.cc
02 May 2025 · 17:40 UTC
ACCEPTED
subject
[ICML 2025] Your paper has been accepted
redacted for author privacy
CVPR
noreply@cvpr.thecvf.com
27 Feb 2026 · 09:02 UTC
ACCEPTED
subject
CVPR 2026 — Notification of Acceptance
redacted for author privacy
AAAI
aaai26-pc@aaai.org
09 Dec 2025 · 22:15 UTC
ACCEPTED
subject
AAAI-26 Paper Decision: ACCEPT
redacted for author privacy

Objections, answered

The honest version.

≥ 40 cited sources

Retraction-checked and deduplicated. Every claim links to its evidence.

An auditable log

Seeds, data versions and commands — hand it to a skeptic and it holds up.

A named reviewer

A real scientist signs the deliverable and stands behind the decision trail.

Work with us

Request access.

We take on a small number of commissions at a time so a scientist can be assigned to each run. Tell us your field and we'll be in touch.